Conv (dilation = 2$) FC = Fully connected = Broadcast over length Conditioner ↔ = Element-wise addition + = Element-wise multiplication Diffusion-step embedding Residual layer ' = 0 Residual layer ' = 1 Residual layer ' = * − 1 ⋯ Skip connections Output ⋯ FC FC FC Conv1×1 Conv1×1 Conv1×1 Conv1×1 Conv1×1 Bi-DilConv-2- ./0 1 ↔ + + + + ⋅ ⋅ tanh 7 = Connect to next residual layer = Input of each residual layer Conv1×1 ReLU swish ReLU swish © 2023 Sakurai & Matsubara Lab. @ Osaka U 14